mirror of
https://github.com/stablyai/orca.git
synced 2026-09-22 00:02:31 +00:00
Jinwoo-H
db356997a9
CI's addendum to round 4's M1: the frame case failed with the fix present, `expected [ Array(1) ] to deeply equal []`, on a slower runner. What scheduled it: `applyFitScale`, through `scheduleDocumentFrame` like every other frame the document asks for — the document has no other rAF call site. It is not an escape from the registry, so the registry is not what changes here. Why it was counted: React unmounts in two steps. The mutation phase detaches the host, and the passive cleanup that calls `dispose` runs after it — about 1 ms later here, 20 to 35 ms later with the CPU throttled 20x, which is the runner shape this failed on. A frame served in that gap runs with a detached container while the document is still live and has not been asked to stop, and nothing could have taken it back: `cancelDocumentFrames` had not been called yet. The oracle judged by the captured container's connectedness, so it read the gap as a leak. It now counts only what runs after the last statement of `dispose`, which is the class coming off the host, observed on the element because React may have detached it already. The same reading fixes the other direction. The precondition is read at that same moment, and the witness is a refit re-armed from a frame of the test's own, so the document is owed a frame at the end of every frame the browser serves and a dispose cannot land where nothing is owed. The single refit the case used before bought one frame, and the retry loop commits on its first attempt whenever the grid still measures. Evidence: with the boundary removed the case reproduces CI's `Array(1)` in two runs of three unthrottled, and in five of five with the CPU throttled 20x, where the detach-to-dispose gap measures 20 to 35 ms; with it, five green runs; with `cancelDocumentFrames` removed, five red runs, all on the leak read and none on the precondition. Claude-Session: https://claude.ai/code/session_01JNnE9qzUZMMnqpZWCqM3nb